Determinism: Why Recompute-Based Recovery Can Break

Lineage-based recovery rests on an assumption so quiet it is easy to miss, and violating it produces some of the most baffling bugs in Spark: recomputation assumes that replaying a transformation produces the same result it did the first time. If a transformation is non-deterministic, that assumption fails, and recovery can silently produce different data than the partition it is meant to replace. Take a transformation that assigns a random value, or one that depends on the current time, or one whose result depends on the order rows happen to arrive in. The first time it runs, it produces one set of values. If a partition is lost and Spark replays the lineage to rebuild it, the non-deterministic step runs again and produces DIFFERENT values. Now the rebuilt partition does not match what th

About This Interactive Section

This section is part of the Lineage as Fault Tolerance lesson on DataDriven, a free data engineering interview prep platform. Each section includes explanations, worked examples, and hands-on code challenges that execute in real time. SQL queries run against a live PostgreSQL database. Python runs in a sandboxed Docker container. Data modeling problems validate against interactive schema canvases. All content is framed around what data engineering interviewers actually test at companies like Meta, Google, Amazon, Netflix, Stripe, and Databricks.

How DataDriven Lessons Work

DataDriven combines four interview rounds (SQL, Python, Data Modeling, Pipeline Architecture) with adaptive difficulty and spaced repetition. Easy problems get harder as you improve. Weak concepts resurface until you master them. Your readiness score tracks progress across every topic interviewers test. Every lesson section ends with problems you solve by writing and running real code, not by picking multiple-choice answers.